Spatio-Temporal Prediction of Surface Remote Sensing Data in Equatorial Pacific Ocean Based on Multi-Element Fusion Network

Abstract

A basic feature of El Niño is an abnormal increase in the surface temperature of the equatorial Pacific Ocean, which can throw ocean–atmosphere interactions out of balance, resulting in heavy rainfall and severe storms. This climate anomaly causes different levels of impacts worldwide, such as causing droughts in some regions and excessive rainfall in others. Therefore, it is important to determine the formation of El Niño by predicting the changes in the sea surface temperature (SST) in the equatorial Pacific Ocean. In this paper, we propose a multi-element fusion network model based on convolutional long short-term memory (ConvLSTM) and an attention mechanism to predict the SST and analyze the effects of different elemental inputs on the model’s prediction performance using the prediction results. The experimental results show that using the sea surface wind (SSW) and sea level anomaly (SLA) as multi-element inputs to predict the SST overcame the shortcomings of the single-element forecast model, and the prediction accuracy of the two-element fusion model was better than that of the three-element fusion model. In the two-element fusion model, using the SSW as an input predicted the SST with a lower prediction error than using the SLA as an input and had better prediction performance compared with other benchmark models. For predicting the SST in the equatorial Pacific Ocean, the monthly average root mean square error (RMSE) was mainly concentrated in the range of 0.4–0.8 °C, and the regions with a larger error dispersion were located in the spatial range of 5° S–5° N and 130° W–90° W, and the monthly average regional RMSE was mainly concentrated in the range of 0.5–1 °C. Finally, we also validated the prediction performance of the model for the SST in El Niño and La Niña years, and the prediction results of the model in La Niña years were better than those in El Niño years.

1. Introduction

The widespread and persistent anomalous warming of the equatorial Pacific Ocean is an El Niño phenomenon [1,2,3,4] that occurs mainly in the eastern and central equatorial Pacific Ocean, which can lead to global climate anomalies [5]. Normally, monsoonal currents in the tropical Pacific region move from the Americas to Asia, keeping the Pacific surface warm and bringing tropical rainfall to the area around Indonesia [6]. However, when the winds and currents are reversed in El Niño years, the heat flows from the equatorial central Pacific to central America, bringing heavy rainfall to the Pacific coastal countries of South America, which leads to frequent flooding [7,8]. This has also caused severe droughts along the coasts of Indonesia and Australia [9,10,11], and global agricultural production has been affected [12,13]. Therefore, predicting the changes in the sea surface temperature (SST) [14,15] in the equatorial Pacific Ocean to predict the formation of El Niño is particularly important to reduce economic and life losses due to marine disasters [16].

There are two main means of acquiring ocean surface data, direct observation and satellite remote sensing [17,18]. Direct observation collects data using buoys, which has the characteristics of accurate sampling, a low cost and real-time uploading, but its coverage is small due to the influence of multiple factors such as variable sea states and instrumentation maintenance. Satellite remote sensing is based on microwave and infrared observation, which has wider coverage and a high resolution [19,20]. The increasing accumulation and refinement of ocean observation data have laid the data foundation for data-driven prediction methods, which are an important means to study and observe the ocean at present.

Internationally, the Niño3.4 SST anomaly index is used to determine the formation of El Niño–Southern Oscillation (ENSO) events. Ren et al. [21] proposed two new indices to determine the type of ENSO event, including the eastern Pacific ENSO (EP) index using SST anomalies in the Niño3 region and the central Pacific ENSO (CP) index using SST anomalies in the Niño4 region. Coupled ocean–atmosphere models [22], coupled global anomalous circulation models [23] and statistical prediction models [24] are generally used to determine the type of ENSO event.

In a study of traditional methods to predict elements of the marine environment [25], Zhang et al. [26] added a variety of marine elements into the regional ocean model system, such as sea winds (SSWs), tides, heat fluxes, water levels, boundary currents, oblique pressure effects, runoff, etc., and performed numerical simulations of the East China Sea [27]. The measured data and model results were consistent with the long-term trend, and the seasonal distribution structure was the same. Although the application of an ocean model avoids the process of obtaining physical states using complex partial differential equations, it is still a challenging task.

In addition, the use of statistical learning models to predict the SST has received extensive attention and research [28,29]. Wu et al. [30] established a nonlinear prediction system for SST anomalies in the tropical Pacific Ocean by using the method of a multilayer perceptron neural network. Ji et al. [31] combined an autoregressive AR model and a Kalman filtering method based on an empirical orthogonal decomposition method and introduced random factors to reconstruct the weekly mean SST, which predicted the SST in the equatorial Pacific, Indian Ocean and Atlantic Ocean. However, this model has a strong dependence on the region and lacks wider universality.

Recently, deep learning networks have also been gradually applied to predict ocean surface remote sensing data [32]. Han et al. [33] used convolutional neural networks (CNNs) to predict the SST, sea surface salinity (SSS) and sea surface height (SSH) in the Pacific Ocean. Various CNNs have also been developed, such as the Video Pixel Network (VPN) with multiple causal convolutional layers [34], the PredCNN network with multiple expanded causal convolutional layers [35], the improved PredCNN network (PredRNN++) [36] and so on. Liu et al. [37] proposed a convolutional long short-term memory neural network (ConvLSTM) that used a CNN to extract relevant features from historical temporal temperature and salt data and then used LSTM networks for SST prediction. These neural networks with CNNs are not suitable for capturing temporal correlations in the data, although they are capable of spatial feature extraction. Xu et al. [38] proposed a regional convolutional long short-term memory theoretical model (RC-LSTM) with a spatio-temporal information processing capability to solve the problem of regional information loss and conducted experiments in coastal China. Gou et al. [39] proposed a generalized deep learning framework for spatio-temporal ocean data (DeepOcean) using the generative module of a multilayer perceptron (MLP) and the predictive module of a multivariate convolutional LSTM (MVC-LSTM) neural network. Wang et al. [40] proposed a multimodal fusion network (MMFnet) for sea level anomaly (SLA) prediction in the South China Sea, using an ocean data assimilation scheme to fuse a convolutional neural network with an ensemble empirical modal decomposition LSTM network (EEMD-LSTM), which had a better forecasting capability for large-radius eddies in distant and coastal regions. These recurrent neural networks are suitable for modeling the temporal correlations of data, but are not suitable for capturing the spatio-temporal correlations in the data. Hou et al. [41] proposed a multi-source spatio-temporal data fusion model (MUST) to predict the short-term SST using Bicubic Convolutional Interpolation (BCI), Spatio-Temporal Dilated ConvLSTM (ST-DC) and Cross-Data Fusion (CDF).

In addition, attention mechanisms are also a hot research topic in prediction networks; they simulate the function of human visual attention and enable the model to focus on the most important parts of the input sequence, which can achieve the transmission of important information, reduce network redundancy and enhance the nonlinear modeling ability of the model network [42]. For example, a neural network model combining ConvLSTM and an attention mechanism has been used to predict the effective lifetime of wind farm facilities [43]. Zhou et al. [44] proposed the dual cross-attention transformer (DCAT) method, which extracts multi-scale features by connecting multiple dual cross-attention blocks and computes cross-attention features for learning integrated cues from low- and high-frequency information on the detection of changes in the image input. Zhou et al. [45] proposed a specific self-attention-based neural network model (3D-Geoformer) to predict temperature anomalies and wind stress anomalies in the upper layer of the three-dimensional ocean. The above attention-based model can acquire new attention features and has better learning effects, but the model lacks the wide applicability and portability needed for different models to adapt to different application fields.

However, the difficulty of marine environmental element prediction is that the ocean dynamics process is complex and nonlinear, especially when El Niño occurs, causing special changes in the SST, which makes prediction more challenging. Through the quantization of Pearson correlation coefficients, we found that there is a certain degree of correlation between marine environmental elements [26], by which the SST is moderately correlated with the SLA and strongly correlated with the SSW. Therefore, whether the fusion of multi-element features achieved by introducing an attention mechanism can improve the prediction accuracy of the SST is the main concern in this paper.

This paper discusses the effects of multi-element inputs and fusion on the prediction performance of the ConvLSTM model and the application of SST prediction models in the equatorial Pacific Ocean, which provides a reference for predicting the formation of El Niño in order to reduce the loss from marine disasters. The details of the study are as follows:

(1)

A multi-element fusion network model based on ConvLSTM and an attention mechanism is proposed, which includes the multi-element input layer, 3D convolutional layer, attention fusion layer, ConvLSTM layer, and single-element prediction output layer. The 3D convolution can better extract the spatio-temporal information of the elements, and the attentional mechanism can fuse the hidden features of multiple elements in the prediction of the SST, and the use of encoding and decoding ConvLSTM enables the spatio-temporal features of the SST to be better learned.

(2)

Prediction accuracy experiments with multiple-element inputs and feature fusion were designed to analyze the effects of different elemental inputs on the prediction performance of the multi-element fusion model using the prediction results. A comparative analysis with other benchmark model prediction results was also conducted to evaluate the model performance according to different indicators.

(3)

The prediction effect of the multi-element fusion network model for the equatorial Pacific Ocean was analyzed, and the prediction ability of the model for the SST in El Niño and La Niña years was verified.

The chapters of this paper are structured as follows: Section 2 describes the data sources and the structure of the multi-element fusion network; Section 3 describes the design of prediction accuracy experiments with multiple-element inputs and feature fusion and analyzes the SST prediction effect for the equatorial Pacific Ocean; Section 4 evaluates the performance of the model in predicting the SST in El Niño and La Niña years; and Section 5 summarizes the conclusions.

2. Methods and Models

2.1. Multi-Element Data Sources

In this paper, the ocean surface remote sensing data used as multi-element inputs were the SST, SLA and SSW. The SST data were the global diurnal satellite observation data provided by the National Oceanic and Atmospheric Administration (NOAA) of the United States, using a new land mask approach with high accuracy, which had a grid resolution of 0.25° and was fully compliant with the GDS2.0 data specification standard [46]. The SLA data were the global diurnal gridded data provided by the Center for Space Research (CNES) of France in the Archiving, Validation and Interpretation of Satellite Oceanographic (Aviso) data, which had a grid resolution of 0.25° and included a wide range of observations [47]. The SSW data were high-resolution hybrid sea wind data from the NOAA, which were generated from a mixture of observations from multiple satellites, reducing data gaps and random errors in sampling [48]. The grid accuracy was the same as for the SST and SLA data.

In this study, the experimental region chosen in which to determine El Niño generation was the east–central Pacific Ocean, and the spatial range of the multi-element ocean surface data used was 5° S–5° N and 180° W–90° W [49,50,51], and the time range was 1999–2018, of which the data from 1999–2014 were used as training and validation data, and the data from 2015–2018 were used as test data. The statistics of the valid grid data for the experimental area are shown in

Table 1. Statistics of grid points in the experimental area (5° S–5° N, 180° W–90° W).

Statistics	SST	SLA	SSW
Total grid points	14,400	14,400	14,400
Invalid grid points (Nan)	2	3	21
Valid grid points	14,398	14,397	14,379

, where Nan indicates land or missing data. Temporally, the total amount of data for 20 years was about 7300 days, and the actual number of missing days was not more than 20 days, with a very low missing rate of about 0.27%. Spatially, the grid size of the spatial points in the experimental area was 40 × 360, and the actual number of missing spatial points among 14,400 spatial points was not more than 21, and the missing rate was also very low, about 0.15%. With a very low missing data rate, the impact on the model performance was limited.

2.2. Methods

2.2.1. ConvLSTM Network

Regarding the problem of spatio-temporal sequence prediction, Shi et al. [52] first proposed a convolutional long short-term memory network (ConvLSTM) for short-term precipitation forecasting in 2015. When performing multi-step forecasting, the ConvLSTM model can fully exploit the high dimensionality of spatio-temporal sequences to effectively capture the spatio-temporal structure and spatial characteristics of the data. Due to the efficient processing of spatio-temporal sequence information by ConvLSTM networks, they are widely used in other research fields, such as image classification [53] and target detection [54]. The structural framework and basic principles of ConvLSTM are the same as those of LSTM, and the structure is shown in Figure 1.

In contrast to LSTM, a distinctive feature of ConvLSTM is that all input cells, state cells, hidden cells, and gating cells are replaced by vectors as three-dimensional tensors, which can be imagined as a series of vectors in a spatial grid. The ConvLSTM model uses a convolution kernel to predict the future state of the grid cells from the adjacent inputs and past states of the space, with the following update mechanism [52]:

Forgotten gate:

$$f_t=\sigma \left(W_{xf}\ast X_t+W_{hf}\ast H_{t-1}+W_{cf}\circ C_{t-1}+b_f\right)$$

(1)

Input gate:

$$i_t=\sigma \left(W_{xi}\ast X_t+W_{hi}\ast H_{t-1}+W_{ci}\circ C_{t-1}+b_i\right)$$

(2)

Status update:

$${\tilde{C}}_t=tanh\left(W_{xc}\ast X_t+W_{hc}\ast H_{t-1}+b_C\right)$$

(3)

$$C_t=f_t\circ C_{t-1}+i_t\circ {\tilde{C}}_t$$

(4)

Output gate:

$$o_t=\sigma \left(W_{xo}\ast X_t+W_{ho}\ast H_{t-1}+W_{co}\circ C_t+b_o\right)$$

(5)

$$H_t=o_t\circ tanh\left(C_t\right)$$

(6)

where the input $X_t$ of the current moment, the hidden $H_{t-1}$ and the state $C_{t-1}$ of the previous moment are the inputs of the ConvLSTM cell, and the outputs are the hidden $H_t$ and the state $C_t$ of the current moment. The state $C_t$ is a “memory canvas” that is updated across timesteps, and the hidden $H_t$ is a “visible output” reflecting the spatio-temporal features at the current moment. The Hadamard product ∘ represents element-by-element multiplication, ∗ denotes the convolution operation, the sigmoid activation function $\sigma$ outputs 0∼1 to control the flow of information, and the hyperbolic tangent activation function tanh outputs −1∼1 to scale the candidate values. $f_t$, $i_t$ and $o_t$ denote the values of the gating unit at the current moment; when the output is close to 0 or 1, it conforms to the physical sense of off and on, and thus acts as a gate control. $W_{x(f,i,c,o)}$, $W_{h(f,i,c,o)}$ and $W_{c(f,i,o)}$ are the weight matrices of the gating unit, and the weights W in the ConvLSTM model are convolution kernels, which capture the local spatial patterns through convolution operations and are used for the extraction of the spatio-temporal features and to control the flow of information. $b_f$, $b_i$, $b_C$ and $b_o$ are the scalar offsets of the gating unit and are used to adjust the initial activation threshold of the gate.

2.2.2. Correlation Analysis of Multiple-Source Marine Environmental Elements

The Pearson correlation coefficient is a linear correlation coefficient that measures the degree of linear correlation between two variables [26]. If there are two marine environmental element variables, X and Y, then the Pearson correlation coefficient of variables X and Y, usually denoted by r, is defined as follows:

$$r={\displaystyle \frac{{\sum }_{i=1}^n(X_i-\overline{X})(Y_i-\overline{Y})}{\sqrt{{\sum }_{i=1}^n{(X_i-\overline{X})}^2}\sqrt{{\sum }_{i=1}^n{(Y_i-\overline{Y})}^2}}}$$

(7)

where n is the sample capacity. The correlation between the two environmental variables and the SLA, SSW and SST is determined by calculating the Pearson correlation coefficient, and the results of the correlation coefficient analysis are shown in Figure 2.

The region in Figure 2a refers to the experimental area (5° S–5° N, 140° W–130° W), and the midpoint in Figure 2b refers to the location (0°, 135° W). From Figure 2a, it can be seen that the regional SST was strongly correlated with the SLA and moderately correlated with the SSW. From Figure 2b, it can be seen that the midpoint SST was strongly correlated with the SLA and weakly correlated with the SSW. Therefore, the SST, SLA and SSW data could be used to constitute the multi-element fusion data used as the input data for the fusion model predicting the SST, which could be used to study the prediction performance of the fusion model under different combinations of multi-element inputs.

2.2.3. Multi-Element Fusion Network Model

Aiming to predict the characteristics of the spatio-temporal sequence of multi-source ocean elements in the equatorial Pacific Ocean, a multi-element fusion method for predicting the spatio-temporal sequence of the regional SST is proposed based on the spatial feature extraction of a 3D CNN and the feature-weighted fusion of an attention mechanism, combinined with the feature of ConvLSTM that can capture the spatio-temporal dependence of medium- and long-term data to realize the spatio-temporal sequence prediction of the SST in the equatorial Pacific Ocean.

In this paper, we introduce an example multi-element fusion network model structure and a multi-element fusion method for predicting the SST using three elements, the SST, SLA and SSW, denoted as the TAWF(sst, sla, ssw) method, which consists of a multi-element input layer, 3D convolutional layer, attention fusion layer, ConvLSTM layer and single-element prediction output layer, as shown in Figure 3. The functional description and parameter settings of the network units used in Figure 3 are shown in

Table 2. Functional description and parameter settings of network units.

Module Units	Functional Description	Parameter Setting
Conv3D	Extracting temporal feature timing information	filters = 40, kernel_size = (3, 3, 3)
Conv3D_A	Extracting temporal feature timing information	filters = 1, kernel_size = (1, 1, 1)
Add	Fuse features of multiple elements so only the amount of information increases, without changing the dimensionality of the image	-
Activation	Perform processing of nonlinear mapping	‘relu’ or ‘sigmoid’
Transpose	Swap different dimensions of the input tensor	perm = [0, 4, 2, 3, 1]
Multiply	Compute the product (element by element) of the input tensor list	-
ConvLSTM2D	Capture long time series dependencies, memorize and store information	filters = 32, kernel_size = (3, 3, 3), recurrent_dropout = 0.2
Conversion	Convert the input shape (timesteps, rows, cols) to (targetsizes, rows, cols)	target size = targetsizes
Dense	Mapping the extracted features to the output through a nonlinear transformation	units = 1

Open-source machine learning framework: TensorFlow-gpu.

. The proposed model was built in PyCharm 2019 software using the Python language, where a TensorFlow-gpu 2.0 and CUDA 10.2 were used for model training and GPU hardware acceleration to increase the run speed and reduce the training time.

The structure of the multi-element fusion network model is described as follows:

(1)

Multi-Element Input Layer:

The types of multi-element input include the SST, SLA and SSW. Firstly, the multi-element data are subjected to a maximum–minimum normalization process to standardize the multi-element data, eliminate the order-of-magnitude differences between various marine environmental element data due to the differences in numerical ranges and units of measurement and scale the data to a reasonable range of intervals so as to improve the training efficiency and make it easier for the model to converge. The normalization process equation is as follows:

$$x_{\mathrm{std}}={\displaystyle \frac{x_i-x_{min}}{x_{max}-x_{min}}}$$

(8)

where $x_{\mathrm{std}}$ is the data in the interval [0, 1] after normalization, $x_{\max }$ and $x_{\min }$ are the maximum and minimum values of the original data in the region and $x_i$ is the original data.

Finally, the normalized data are fed into the 3D convolutional layer. The input shapes of the SST, SLA and SSW are denoted by (None, timesteps, rows, cols, 1), where None, timesteps, rows, cols and 1 denote the number of samples, the input timestep, the rows of the data matrix, the columns of the data matrix and the number of channels.

(2)

3D Convolutional Layer:

The function of the 3D convolutional layer is to extract the hidden temporal and spatial information of the elements by using Conv3D units. A Conv3D unit is a 3-dimensional convolution process, which completes the convolution kernels and the element values at the corresponding position obtained by first multiplying the product before summation, by moving the 3 × 3 × 3 convolution kernels in the 2-dimensional image plane and the 3rd-dimensional depth direction. Table 2 shows the parameter settings of Conv3D, and a 3-dimensional convolution operation is performed for each of the 3 elements, using 40 filters, so the data shape is (None, timesteps, rows, cols, 40).

(3)

Attention Fusion Layer:

The attention fusion layer mainly utilizes the attention mechanism to fuse the correlated features among the three multi-elements of the SST, SLA and SSW and reassign feature weights to the predicted SST elements to generate new weighted features. In this case, the attention mechanism assigns different weights to all input feature sequences, and the different feature attention weights are jointly determined by the degree of correlation between the model input features and the network output. During the training process, the attention mechanism continuously matches important information with higher weights and filters out invalid information that can be ignored with lower weights. As the training times accumulate, the attention weight matrix used for information matching is continuously optimized, and the trained neural network parameters become more usable.

This workflow is divided into two parts, as shown in Figure 4. In one part, hidden information about the SST, $HSS{T}_i$, is obtained using the Transpose unit. In the other part, the feature weight ${\beta }_i$ of multiple elements (SST, SLA and SSW) is obtained through the continuous use of the Add, ReLU, Conv3D_A, Sigmoid and Transpose units, where Table 2 shows the functional description and parameter settings of the module units. Finally, ${\beta }_i$ and $HSS{T}_i$ are multiplied to obtain the new feature $SS{T}_i$’, the data shape of which is (None, 40, rows, cols, timesteps).

(4)

ConvLSTM Layer:

The function of the ConvLSTM layer is to further extract features, and it mainly captures long time sequence dependencies, augments the time taken to generate target lengths and remembers and stores information. The structure of the ConvLSTM layer is shown in Figure 3, and the functional description and parameter settings are shown in Table 2. The output is obtained through the continuous application of the Transpose, 2 ConvLSTM 2D, Conversion and 2 ConvLSTM 2D units, and the data shape is (None, targetsizes, rows, cols, 32).

(5)

Single-Element Prediction Output Layer:

The fully connected layer maps the extracted features to the output space using a nonlinear transformation and then visualizes the prediction results through inverse normalization. The inverse normalization equation is as follows:

$$x_i=x_{min}+x_{\mathrm{std}}·\left(x_{max}-x_{min}\right)$$

(9)

where the symbolic interpretation is the same as in Equation (7), and $x_i$ is the visualization of the true prediction at each position, i, in the [rows, cols] region. The shape of the output prediction result is represented by (1, targetsizes, rows, cols, 1), where targetsizes represent the number of days directly predicted for the output.

This structure of a multi-element fusion network model based on ConvLSTM and an attention mechanism was selected for the experiments in this paper. In the 1999–2014 data set, 80% and 20% of the data were used as the training set and validation set, respectively. The data from 2015–2018 were used as the test set. A control variable approach was used to optimize the model, and the training parameters of the model were determined as follows: optimizer = ‘adam’, learning_rate = 0.001, loss = ‘mse’, metrics = [‘cosine_similarity’] and epochs = 100.

2.3. Evaluation Indicators

In this study, the Bias, root mean square error (RMSE), prediction accuracy (PACC) and conditional percentage (${\mathrm{P}}_{(\backsim )}$) were used to evaluate the prediction performance of the model.

$$Bias=\left|Y_{pre}-Y_{obs}\right|$$

(10)

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}{\sum }_{i=1}^N{\left(Y_{pre}-Y_{obs}\right)}^2}$$

(11)

$$PACC=1-{\displaystyle \frac{1}{N}}{\sum }_{i=1}^N\left({\displaystyle \frac{\left|Y_{pre}-Y_{obs}\right|}{Y_{obs}}}\right)$$

(12)

$$P_{(\backsim )}={\displaystyle \frac{N_{(\backsim )}}{N}}\times 100\%$$

(13)

where $Y_{obs}$ and $Y_{pre}$ are the observed and predicted values of the sea surface elements, and $(\backsim )$ denotes the conditions satisfied. It should be noted that when N is the total number of grid points, the regional indicators Reg_RMSE and Reg_PACC are obtained. However, when N is the length of time, the error indicators Poi_RMSE and Poi_PACC are obtained for each grid point. The smaller the RMSE, the more accurate the prediction.

3. Results

3.1. Prediction Accuracy Experiments with Multi-Element Input and Feature Fusion

Within the experimental region described in Section 2.1, we chose a region with a spatial extent of 5° S–5° N and 140° W–130° W for the experiments and determined the parameters as follows: rows = 40 and cols = 40.

In the validation experiments, our multi-element network fusion model had different structures depending on the input elements, such as the following:

(1)

TF(sst): Input the SST and predict the output SST.

(2)

TAF(sst, sla): Input the SST and SLA and predict the output SST.

(3)

TWF(sst, ssw): Input the SST and SSW and predict the output SST.

(4)

TAWF(sst, sla, ssw): Input the SST, SLA and SSW and predict the output SST.

3.1.1. Experiments on Analyzing Historical Input Lengths and Predicted Output Lengths

In order to analyze the effect of historical inputs and target outputs on the performance of predictive models, we chose the TF(sst) model for our experiments, and the input and output parameters of the model were determined as follows: timesteps = {5, 10, 15, 20, 30, 40}, targetsizes = {5, 10, 15}, rows = 40 and cols = 40. Twelve prediction experiments were performed at 12 time points, the Reg_RMSE and Reg_PACC of each prediction were counted and the average Reg_RMSE and average Reg_PACC of the twelve experiments were calculated;

Table 3. The average Reg_RMSE and average Reg_PACC of the TF(sst) model under different timesteps and targetsizes (5° S–5° N, 140° W–130° W).

Parameters		Day 1		5-Day Average		10-Day Average		15-Day Average
targetsizes	timesteps	Reg_RMSE	Reg_PACC	Reg_RMSE	Reg_PACC	Reg_RMSE	Reg_PACC	Reg_RMSE	Reg_PACC
5	5	0.5933	0.9793	0.6419	0.9780	-	-	-	-
	10	0.4018	0.9876	0.4793	0.9847	-	-	-	-
	15	0.4341	0.9860	0.5322	0.9827	-	-	-	-
	20	0.5271	0.9844	0.5437	0.9827	-	-	-	-
10	10	0.5460	0.9827	0.5620	0.9818	0.6581	0.9782	-	-
	15	0.5104	0.9841	0.5387	0.9827	0.6305	0.9793	-	-
	20	0.4248	0.9874	0.4608	0.9854	0.5547	0.9818	-	-
	30	0.5028	0.9842	0.5335	0.9831	0.6179	0.9800	-	-
	40	0.4418	0.9861	0.5242	0.9831	0.6286	0.9793	-	-
15	15	0.5410	0.9832	0.5480	0.9824	0.6288	0.9793	0.6672	0.9779
	20	0.5620	0.9824	0.5914	0.9806	0.6782	0.9772	0.7231	0.9755
	30	0.4564	0.9865	0.4673	0.9854	0.5403	0.9828	0.5752	0.9816
	40	0.8976	0.9692	0.8452	0.9712	0.8036	0.9730	0.7769	0.9739

shows the average Reg_RMSE and average Reg_PACC of the TF(sst) model under different timesteps and targetsizes. Due to the chance and uncertainty of the prediction experiments, the prediction error may have been small in one time period, but may have been larger in another time period. Therefore, the twelve time points were selected in such a way that only one time point was selected uniformly at the beginning of each month over 12 consecutive months, for example, the first day of each month.

In Table 3, we can see that when targetsizes = 5, in timesteps = {5, 10, 15, 20}, timesteps = 10 had a smaller Reg_RMSE on Day 1 and for the 5-day average; when targetsizes = 10, in timesteps = {10, 15, 20, 30, 40}, timesteps = 20 had the smallest Reg_RMSE and the largest Reg_PACC; and when targetsizes = 15, in timesteps = {15, 20, 30, 40}, timesteps = 30 had the best predictive performance. By comparing Reg_RMSE and Reg_PACC based on the 5-day average, we found that Reg_RMSE was smallest when targetsizes = 10 and timesteps = 20, followed by when targetsizes = 15 and timesteps = 30. However, comparing Reg_RMSE and Reg_PACC based on the 10-day average, the minimum value of Reg_RMSE was 0.5403 and the maximum value of Reg_PACC was 0.9828 when targetizes = 15 and timesteps = 30. At the same time, forecasting 15 days is more advantageous for applications with longer forecast ranges. Therefore, the use of 30 days of data as the input to directly predict the SST for the next 15 days resulted from the optimal set of experiments.

3.1.2. Experiments on Analyzing Multi-Element Fusion Methods Under Different Element Inputs

(i) Analysis of one prediction experiment

A set of data in the test set was selected for one prediction experiment, using 30 days of multi-element data as the input and directly predicting the output SST for the next 15 days. Figure 5 shows the prediction results for different model structures obtained from a one-time experiment, where ‘real’ denotes the real SST in the region. Through direct observation, we found that the results for multi-element fusion network models with different structures were consistent with the magnitude of temperature changes in the real SST.

Figure 6 shows the prediction Bias for different model structures in one experiment, with the Bias calculated using Equation (9). All models in Figure 6 had the smallest prediction Bias on Day 1, and the error started to increase with time. Through direct observation, we found that the prediction Bias was larger in the region with a latitude range of 0–5° N compared to the region with a latitude range of 0–5° S.

We calculated the RMSE and PACC of the selected region in Figure 6 and obtained the Reg_RMSE and Reg_PACC for different model structures from one experiment as shown in

Table 4. Reg_RMSE and Reg_PACC for different model structures obtained from one experiment (5° S–5° N, 140–130° W).

Indicators	Models	Day 1	Day 3	Day 5	Day 7	Day 9	Day 11	Day 13	Day 15	Ave-Day
Reg_RMSE	TF(sst)	0.5425	0.6359	0.8958	0.8509	0.9445	0.9868	0.9833	0.7932	0.8432
	TAF(sst, sla)	0.2274	0.3126	0.5501	0.5289	0.5941	0.6452	0.6474	0.5180	0.5132
	TWF(sst, ssw)	0.2959	0.3066	0.4267	0.4038	0.4460	0.5148	0.5488	0.5129	0.4316
	TAWF(sst,sla,ssw)	0.2576	0.4225	0.7193	0.7039	0.8002	0.8425	0.8436	0.6736	0.6746
Reg_PACC	TF(sst)	0.9802	0.9776	0.9687	0.9706	0.9664	0.9657	0.9673	0.9746	0.9708
	TAF(sst, sla)	0.9937	0.9901	0.9828	0.9829	0.9804	0.9788	0.9796	0.9842	0.9836
	TWF(sst, ssw)	0.9913	0.9905	0.9870	0.9873	0.9863	0.9840	0.9832	0.9847	0.9868
	TAWF(sst,sla,ssw)	0.9924	0.9861	0.9755	0.9757	0.9713	0.9706	0.9719	0.9786	0.9771

, where Ave-day denotes the 15-day average. The Ave-day Reg_RMSE results, shown in Table 4, were as follows: TF(sst) > TAWF(sst, sla, ssw) > TAF(sst, sla) > TWF(sst, ssw). This indicates that increasing the SLA or SSW could reduce the Reg_RMSE of the model and improve the prediction accuracy, and the two-element fusion model performed better at prediction than the three-element fusion model. The TAF(sst, sla) method had a lower Reg_RMSE and slightly better prediction performance than the TWF(sst, ssw) method on Day 1, verifying that the regional correlation of the SST with the SLA, described in Section 2.2.2, was slightly higher than that with the SSW, mainly on Day 1, but the two-element fusion prediction method for the SST and SSW was superior regarding 15-day predictions except on Day 1. Therefore, the TWF(sst, ssw) method with multi-step prediction for 15 days had the best overall prediction accuracy, especially regarding medium-term prediction.

(ii) Analysis of multiple prediction experiments

In order to evaluate the prediction performance of the model, we selected a period of 600 days from the test set and made predictions for each day, repeating the experiment a total of 600 times.

Figure 7 shows the Poi_RMSE at each location for the different model structures obtained from 600 experiments. In Figure 7, it can be seen that the Poi_RMSE for the different models was very small on Day 1 and started to increase with time, but the TF(sst) model performed the worst in terms of its overall prediction performance, which indirectly indicates that using multiple elements as model inputs can reduce the prediction error to some extent. Meanwhile, we found that the location of the largest Poi_RMSE on the 15th day was mainly centered around a latitude of 0°, which may have been due to the large fluctuation in the SST changes in the equatorial region.

(iii) Performance analysis of space points in region

Several specific locations in the 5° S–5° N and 140° W–130° W region in Figure 7 were selected for which the results were analyzed individually, and the specific locations selected in this study were as follows:

(1)

Location A: In Figure 5, the location with the highest real SST in the real 1-day data.

(2)

Location B: In Figure 5, the location with the lowest real SST in the real 1-day data.

(3)

Location C: In Figure 7, the location with the smallest Poi_RMSE in the 1-day TF(sst) model.

(4)

Location D: In Figure 7, the location with the smallest Poi_RMSE in the 1-day TWF(sst, ssw) model.

Table 5. Ave-day Poi_RMSE and Poi_PACC at the four specific locations for the different model structures obtained from 600 experiments.

Models	Location A		Location B		Location C		Location D
	Poi_RMSE	Poi_PACC	Poi_RMSE	Poi_PACC	Poi_RMSE	Poi_PACC	Poi_RMSE	Poi_PACC
TF(sst)	0.5717	0.9829	0.8618	0.9746	0.3725	0.9887	0.3908	0.9880
TAF(sst, sla)	0.3570	0.9903	0.7760	0.9760	0.3252	0.9904	0.3030	0.9913
TWF(sst, ssw)	0.3764	0.9896	0.7448	0.9776	0.3019	0.9911	0.2732	0.9921
TAWF(sst,sla,ssw)	0.4316	0.9878	0.7769	0.9765	0.3572	0.9891	0.3377	0.9895

shows the Ave-day Poi_RMSE and Poi_PACC at the four specific locations for the different model structures obtained from 600 experiments. By comparing the Poi_RMSE and Poi_PACC, we found that the Poi_RMSE of the TWF(sst, ssw) model at location B, location C and location D was 0.7448, 0.3019 and 0.2732, respectively, and the Poi_RMSE of the TWF(sst, ssw) model was the smallest in all three locations, while the Poi_PACC was the largest in all of them. Meanwhile, the Poi_RMSE of the TAF(sst, sla) model was the smallest, with a value of 0.3570 at location A. In order to more conveniently observe the Poi_RMSE for the different model structures at the four specific locations, we compared the Poi_RMSE values in Table 5 by plotting the histograms, as shown in Figure 8.

In terms of the overall performance, the TWF(sst, ssw) model had a smaller prediction error and higher prediction accuracy.

3.2. Comparative Analysis with Other Model Prediction Results

From the experiments described in Section 3.1, we concluded that the TWF(sst, ssw) model of the multi-element fusion network model had a smaller prediction error and higher prediction accuracy. Among the many deep learning methods introduced, we chose two classical benchmark models based on LSTM [38] and ConvLSTM [37] and the derived attention mechanism models ATT-LSTM and ATT-ConvLSTM [43]. In the validation experiments described in Section 3.2, we compared the proposed TWF(sst, ssw) model with the above four benchmark models to analyze the prediction performance of the models. The detailed model structures are described as follows:

(1)

Multi-Channel LSTM [38]: The encoder and decoder are built using two LSTM layers, respectively, and depending on the elements of the multi-channel inputs, we can obtain LSTM and Multi-LSTM.

(2)

Multi-Channel ATT-LSTM: By introducing an attention mechanism to a multi-channel LSTM foundation, the corresponding ATT-LSTM and Multi-ATT-LSTM models can be obtained.

(3)

ConvLSTM [37]: Using a single-element model with only an SST input and two ConvLSTM2D layers to build an encoder and decoder, respectively, we can obtain ConvLSTM. We can also obtain CNN+ConvLSTM by adding a CNN layer.

(4)

ATT-ConvLSTM [43]: A channel attention mechanism is introduced to fuse multi-element inputs based on ConvLSTM to obtain ATT-ConvLSTM-Fusion. It is worth noting that the channel attention mechanism is different from the attention mechanism used in the TWF(sst, ssw) model proposed in this paper.

(5)

LICOM [55]: The global ocean circulation model developed by the Institute of Atmospheric Physics of the Chinese Academy of Sciences is numerically solved using a finite difference method under the given initial and boundary values and is applied to short- and medium-term marine environmental forecasting.

Referring to the experimental procedure in Section 3.1, we selected a period of 24 months, randomly chose 2 time points for each month and obtained 48 time points in total. Prediction experiments were conducted for the 48 randomly selected time points, and the average of the 48 experiments was calculated to obtain the average Reg_RMSE and average Reg_PACC.

Table 6. Comparison of the average Reg_RMSE and average Reg_PACC results for other models.

Indicators	Models	Day 1	Day 3	Day 5	Day 7	Day 9	Day 11	Day 13	Day 15	Ave-Day
Reg_RMSE	LICOM	0.5220	0.5590	0.6000	0.6470	-	-	-	-	0.5901
	LSTM	0.6602	0.6479	0.7320	0.7694	0.7703	0.7824	0.8033	0.8193	0.7478
	ATT-LSTM	0.5034	0.5774	0.6222	0.6817	0.6804	0.6857	0.6977	0.7099	0.6493
	ConvLSTM	0.4095	0.4401	0.5263	0.5955	0.6234	0.6563	0.6880	0.7062	0.5798
	CNN+ConvLSTM	0.4501	0.4841	0.5663	0.6342	0.6652	0.6963	0.7240	0.7406	0.6200
	Multi-LSTM	0.5024	0.5941	0.6340	0.6831	0.6928	0.6992	0.7131	0.7264	0.6594
	Multi-ATT-LSTM	0.5225	0.5769	0.6626	0.6985	0.6941	0.7095	0.7320	0.7458	0.6727
	ATT-ConvLSTM-Fusion	0.4072	0.4752	0.5294	0.5805	0.6118	0.6540	0.6909	0.7033	0.5831
	TWF(sst, ssw)	0.3563	0.4026	0.5005	0.5742	0.6052	0.6395	0.6735	0.6950	0.5561
Reg_PACC	LSTM	0.9782	0.9794	0.9765	0.9752	0.9751	0.9748	0.9740	0.9735	0.9759
	ATT-LSTM	0.9843	0.9817	0.9805	0.9782	0.9783	0.9783	0.9778	0.9773	0.9794
	ConvLSTM	0.9883	0.9867	0.9840	0.9817	0.9808	0.9797	0.9785	0.9778	0.9822
	CNN+ConvLSTM	0.9854	0.9847	0.9822	0.9799	0.9792	0.9781	0.9772	0.9765	0.9804
	Multi-LSTM	0.9845	0.9809	0.9798	0.9780	0.9778	0.9777	0.9771	0.9766	0.9789
	Multi-ATT-LSTM	0.9836	0.9819	0.9791	0.9778	0.9779	0.9775	0.9766	0.9762	0.9787
	ATT-ConvLSTM-Fusion	0.9876	0.9854	0.9838	0.9821	0.9812	0.9797	0.9784	0.9779	0.9820
	TWF(sst, ssw)	0.9901	0.9879	0.9848	0.9823	0.9813	0.9802	0.9789	0.9781	0.9829

shows a comparison of the average Reg_RMSE and average Reg_PACC results for other models. From Table 6, it can be seen that the average Ave-day Reg_RMSE and average Ave-day Reg_PACC of the TWF(sst, ssw) model were 0.5561 and 0.9829, respectively. We can see that among all the models, the TWF(sst, ssw) model had the smallest Reg_RMSE and the highest Reg_PACC on every day. Statistically, the results show that the TWF(sst, ssw) model reduced the Reg_RMSE by about 22.41% regarding the 7-day average compared to the traditional ocean circulation LICOM model, and the TWF(sst, ssw) model reduced the Reg_RMSE by about 8.17–35.71% regarding the 7-day average and 3.40–25.33% regarding the 15-day average compared to other deep learning models. The advantages of the TWF(sst, ssw) model for the regional prediction of the SST have been verified.

In order to more visually compare the average Reg_RMSE results of the other models, we provide a comparison of the Reg_RMSE results of the other models on every day, as shown in Figure 9. Through direct observation, we found that ConvLSTM, CNN+ConvLSTM, ATT-ConvLSTM-Fusion and TWF(sst, ssw) had a smaller Reg_RMSE than the other models when directly predicting the SST at Day 15, which was especially pronounced from Day 1 to Day 10. Therefore, we found that adding convolutions or ConvLSTM when building a model could effectively reduce the regional prediction error and improve the regional prediction accuracy, as our proposed TWF(sst, ssw) model had the best prediction effect.

Figure 10 shows the Poi_RMSE at each location for the other prediction models obtained from 48 experiments, with darker colors indicating a larger Poi_RMSE. Through observation, we could see that the Poi_RMSE of ConvLSTM, CNN+ConvLSTM, ATT-ConvLSTM-Fusion and TWF(sst, ssw) had a significantly smaller Poi_RMSE on Day 1, and the Poi_RMSE increased gradually as the number of prediction days increased. The Poi_RMSE of the TWF(sst, ssw) model did not exceed 0.7 °C in 5 days and 1.1 °C in 15 days.

3.3. Prediction Experiment for Equatorial Pacific Ocean

Using the TWF(sst, ssw) model with the best prediction performance, described in Section 3.2, prediction experiments were performed on the experimental region described in Section 2.1, which spatially ranged from 5° S–5° N and 180° W–90° W. The input and output parameters of the model were determined as follows: timesteps = 10, target sizes = 3, rows = 40 and cols = 360. It should be noted that the size of the selected region was 40 × 360, and it consisted of nine 40 × 40 grids.

3.3.1. Prediction Experiments for the Equatorial Pacific Ocean in El Niño Years

The El Niño event occurred between October 2014 and April 2016, and we chose to test it during a 12-month period from July 2015 to June 2016. The TWF(sst, ssw) model was used to directly output the 15-day SST, expand the forecast range with two 15-day forecasts, realize one 30-day forecast per month, study the feasibility of the model’s predictions in the equatorial Pacific Ocean and investigate the applicability of the model using the spatial distribution of the Bias and Poi_RMSE for each month, shown in Figure 11.

Figure 11a shows the spatial distribution of the monthly mean SST predicted by the TWF(sst, ssw) model over a 12-month period in El Niño years. In the monthly prediction results, the SST characteristics in the equatorial Pacific Ocean were clearly visible, with the SST starting to decrease along the longitude of 180° to 90° W with a gradient, while the closer to the equatorial latitude near 0° the SST was predicted, the lower it became.

We evaluated the TWF(sst, ssw) model using the Bias and Poi_RMSE. Figure 11b shows the spatial distribution of the monthly mean SST prediction Bias for the TWF(sst, ssw) model over a 12-month period in El Niño years, where the prediction Bias indicates the error between the prediction and the real SST. It was found that the prediction Bias was mostly less than 0.9 °C, and the prediction Bias was relatively large in the months from November 2015 to June 2016, especially in June 2016, when the Bias was the largest, and it was mainly concentrated in the sea area in the longitude range of 130–90° W. As can be seen from Figure 11c, the monthly mean SST prediction Poi_RMSE during the 12-month period was mostly less than 1.2 °C, and similarly, there was a large Poi_RMSE in June 2016 in the sea area of the longitude range of 130–90° W. In general, the regions with a larger error dispersion over the 12-month period were located in the 5° S–5° N and 130–90° W spatial range.

We calculated the percentage of the mean Poi_RMSE for each month, shown in Figure 11c, within the ranges of Poi_RMSE < 0.4, 0.4 < Poi_RMSE < 0.8, 0.8 < Poi_RMSE < 1.2, 1.2 < Poi_RMSE < 1.6, 1.6 < Poi_RMSE < 2.0 and Poi_RMSE > 2.0, respectively, as shown in

Table 7. ${\mathrm{P}}_{(\backsim )}$ for different Poi_RMSE range intervals over a 12-month period in El Niño years.

${\mathbf{P}}_{\mathbf{(}\mathbf{\backsim }\mathbf{)}}$	Jul. 2015	Aug. 2015	Sep. 2015	Oct. 2015	Nov. 2015	Dec. 2015	Jan. 2016	Feb. 2016	Mar. 2016	Apr. 2016	May 2016	Jun. 2016
${\mathrm{P}}_{(Poi\_RMSE<0.4)}$	30.38%	38.72%	36.15%	17.07%	16.15%	23.53%	12.13%	19.13%	16.94%	8.98%	13.91%	8.91%
${\mathrm{P}}_{(0.4<Poi\_RMSE<0.8)}$	61.31%	50.36%	47.99%	53.28%	38.33%	56.18%	51.84%	48.76%	42.07%	35.24%	43.85%	41.92%
${\mathrm{P}}_{(0.8<Poi\_RMSE<1.2)}$	7.33%	8.05%	13.47%	26.02%	29.56%	18.05%	29.77%	21.81%	21.53%	36.28%	22.31%	28.16%
${\mathrm{P}}_{(1.2<Poi\_RMSE<1.6)}$	0.94%	1.79%	1.92%	3.51%	14.26%	2.19%	5.83%	8.69%	15.58%	17.92%	15.40%	13.30%
${\mathrm{P}}_{(1.6<Poi\_RMSE<2.0)}$	0.05%	0.72%	0.42%	0.11%	1.71%	0.06%	0.40%	1.60%	3.64%	1.29%	3.89%	4.51%
${\mathrm{P}}_{(Poi\_RMSE>2.0)}$	-	0.36%	0.03%	-	-	-	0.03%	0.01%	0.24%	0.29%	0.63%	3.19%

. It was found that ${\mathrm{P}}_{(0.4<Poi\_RMSE<0.8)}$ was the largest in all months except April 2016. In general, the monthly mean Poi_RMSE was mainly concentrated in the range of 0.4–0.8 °C.

In order to quantify the predictive ability of the TWF(sst, ssw) model for each month, we provide the monthly average Reg_RMSE variation curves for each month in El Niño years, as shown in Figure 12. Through observation, we found that the monthly average Reg_RMSE was mainly concentrated in the range of 0.5–1 °C with an increasing trend and was the largest in June 2016. The overall prediction performance of the TWF(sst, ssw) model in the equatorial Pacific Ocean remained relatively perfect during El Niño years.

3.3.2. Prediction Experiments for the Equatorial Pacific Ocean in La Niña Years

The La Niña event occurred between October 2017 and March 2018, and we chose to test it during a 12-month period from July 2017 to June 2018. Using the TWF(sst, ssw) model and data analysis methods described in Section 3.3.1, the spatial distribution of the monthly mean sea surface temperature prediction Bias and Poi_RMSE was similarly obtained, as shown in Figure 13.

In the monthly prediction results, the SST gradient boundary is clearer in Figure 13a compared to Figure 11a. The prediction Bias, shown in Figure 13b, was mostly less than 0.8 °C, and the monthly mean SST prediction Poi_RMSE, shown in Figure 13c, was mostly less than 1.2 °C. The locations with a large prediction Bias and Poi_RMSE were mainly concentrated in the sea area within the longitude range of 110–90° W.

In

Table 8. ${\mathrm{P}}_{(\backsim )}$ for different Poi_RMSE range intervals over a 12-month period in La Niña years.

${\mathbf{P}}_{(\backsim )}$	Jul. 2017	Aug. 2017	Sep. 2017	Oct. 2017	Nov. 2017	Dec. 2017	Jan. 2018	Feb. 2018	Mar. 2018	Apr. 2018	May 2018	Jun. 2018
${\mathrm{P}}_{(Poi\_RMSE<0.4)}$	10.28%	21.31%	10.78%	8.57%	6.16%	8.77%	5.93%	23.78%	24.49%	28.24%	36.28%	38.92%
${\mathrm{P}}_{(0.4<Poi\_RMSE<0.8)}$	41.99%	36.11%	47.35%	41.49%	36.38%	50.96%	50.05%	61.22%	46.92%	63.26%	54.65%	50.47%
${\mathrm{P}}_{(0.8<Poi\_RMSE<1.2)}$	39.24%	23.94%	23.50%	30.31%	28.45%	28.22%	32.28%	10.58%	18.40%	6.56%	8.22%	7.40%
${\mathrm{P}}_{(1.2<Poi\_RMSE<1.6)}$	7.65%	13.69%	12.17%	14.56%	19.23%	9.74%	9.41%	3.86%	6.97%	1.20%	0.66%	2.56%
${\mathrm{P}}_{(1.6<Poi\_RMSE<2.0)}$	0.77%	3.95%	5.56%	3.89%	7.95%	1.99%	1.95%	0.47%	2.36%	0.48%	0.11%	0.64%
${\mathrm{P}}_{(Poi\_RMSE>2.0)}$	0.08%	1.00%	0.63%	1.18%	1.83%	0.32%	0.38%	0.08%	0.86%	0.27%	0.08%	0.01%

, it can be seen that ${\mathrm{P}}_{(0.4<Poi\_RMSE<0.8)}$ was the largest in each of these months, followed by ${\mathrm{P}}_{(0.8<Poi\_RMSE<1.2)}$, and more exceptionally, ${\mathrm{P}}_{(Poi\_RMSE<0.4)}$ was larger than ${\mathrm{P}}_{(0.8<Poi\_RMSE<1.2)}$ from February to June 2016. In general, the monthly mean Poi_RMSE was mainly concentrated in the range of 0.4–0.8 °C.

Figure 14 shows the monthly average Reg_RMSE variation curves for each month in La Niña years. Through observation, we found that the monthly average Reg_RMSE was mainly concentrated in the range of 0.5–1 °C with a decreasing trend and was the largest in November 2017. The overall prediction performance of the TWF(sst, ssw) model in the equatorial Pacific Ocean remained relatively perfect during La Niña years.

4. Discussion

In Section 3.3, we discussed the SST in the equatorial Pacific Ocean over a 12-month period in El Niño and La Niña years. Here, we focus on the monthly average anomalous variations in the SST in the study of El Niño and La Niña events.

An anomaly in the SST is a difference between the SST and the multi-year average SST. We obtained the multi-year average SST by averaging the SST over the time period from 1999 to 2014. Figure 15 shows the monthly average anomalies in the SST over a 12-month period in El Niño years. We found that the monthly average anomalies in the SST were more variable, reaching a maximum of 6 °C and a minimum of −2 °C. Similarly, the anomalies in the SST for the period from January 2016 to May 2016 were more drastically variable. Figure 16 shows the monthly average anomalies in the SST over a 12-month period in La Niña years. We found that the monthly average anomalies in the SST were more variable, reaching a maximum of 4 °C and a minimum of −5 °C.

The U.S. defines an El Niño event as a 3-month sliding average of SST anomalies reaching 0.5 °C or more in the Niño3.4 region and a La Niña event as a 3-month sliding average of SST anomalies less than or equal to −0.5 °C in the Niño3.4 region, where the spatial extent of the Niño3.4 region is 5° N–5° S and 170–120° W [56,57]. We calculated the regional monthly average SST anomalies for the Niño3.4 region, shown in Figure 15 and Figure 16, respectively, and obtained the regional monthly average anomaly variation curves, as shown in Figure 17 and Figure 18.

In Figure 17, it can be seen that the Niño3.4 region’s monthly average anomaly was more than 0.5 °C in each month from July 2015 to June 2016, and the SST anomaly exceeded 1.5 °C in 10 consecutive months, peaking in November 2015 with a regional monthly average anomaly of 1.87 °C. Therefore, we determined that the 3-month sliding average of SST anomalies exceeded 0.5 °C, and an El Niño event was judged to have occurred. The temporal evolution of the El Niño 3.4 index predicted by Zhou et al. [45] was very similar to the observations, albeit with a weaker amplitude, with the peak intensity reaching a maximum in November 2015, generating an El Niño 3.4 SST anomaly intensity of about 2.8 °C, compared to an actual observed anomaly of about 3.0 °C. In contrast, our predicted SST anomaly for November 2015 was 1.87 °C, and the actual observed anomaly was about 2.74 °C. Although there were differences, the overall trend was the same. In Figure 17, it can be seen that the real monthly average anomalies and the monthly average anomalies predicted by the TWF(sst, ssw) model had the same fluctuation trend with a high degree of fit, and the largest difference was in November 2015, which indicates that the TWF(sst, ssw) model had better prediction performance in El Niño years.

In Figure 18, it can be seen that the Niño3.4 region’s monthly average anomaly was less than −0.5 °C in each month from November 2017 to March 2018. Therefore, we determined that the 3-month sliding average of SST anomalies was less than −0.5 °C, and a La Niña event was judged to have occurred. In Figure 18, the black real results almost overlap with the blue TWF(sst, ssw) results, indicating that the monthly average anomalies predicted by the TWF(sst, ssw) model were very close to the real monthly average anomalies and that the TWF(sst, ssw) model had better prediction performance in La Niña years.

Meanwhile, the results were consistent with the actual El Niño and La Niña events monitored by the World Meteorological Organization (WMO) [58] and the National Climate Center of China (NCC) [59] from 2015 to 2016 and 2017 to 2018.

5. Conclusions

In this paper, we focused on a multi-element fusion network model based on ConvLSTM and an attention mechanism, using a Conv3D layer to extract the temporal and spatial features of the elements and fusing the features of multiple elements with an attention fusion layer, so as to improve the prediction accuracy of the SST. Firstly, a spatial range of 5° S–5° N and 140–130° W was selected as the study area, and the multi-element fusion network model was evaluated using the RMSE and PACC. Then, the prediction models were compared with other benchmark models (LSTM, ATT-LSTM, ConvLSTM and ATT-ConvLSTM). Finally, an area of the equatorial Pacific Ocean with a spatial range of 5° S–5° N and 180–90° W was selected as the study area to analyze the prediction effect in the equatorial Pacific region and to validate the applicability and prediction feasibility of the model in El Niño and La Niña years by examining the monthly average anomalous variations in the SST. The main conclusions drawn were as follows.

(1)

The accuracy of the model was evaluated using the RMSE and PACC. The TWF(sst, ssw) model had the smallest Reg_RMSE.mean of 0.4748 °C, and the results for the Reg_RMSE.mean were as follows: TF(sst) > TAWF(sst, sla, ssw) > TAF(sst, sla) > TWF(sst, ssw). The prediction accuracy of the model was affected by the multi-element inputs. The two-element fusion model using the SSW and SLA as the multi-element inputs had smaller prediction errors than the SST single-element model and even smaller prediction errors than the three-element fusion model. Here, the TWF(sst, ssw) model had the best prediction ability.

(2)

The TWF(sst, ssw) model was compared with other benchmark models. The Poi_RMSE in the region with a latitude range of 0–5° N was larger than the Poi_RMSE in the region with a latitude range of 0–5° S among all the models. The TWF(sst, ssw) model had the smallest Reg_RMSE and the highest Reg_PACC on every day. Statistically, this indicates the superiority of the TWF(sst, ssw) model. Meanwhile, adding convolutions or ConvLSTM when building the model could effectively reduce the regional prediction error and improve the regional prediction accuracy.

(3)

In the equatorial Pacific during El Niño years, most of the prediction Bias values for the TWF(sst, ssw) model were less than 0.9 °C, and the monthly mean Poi_RMSE was less than 1.2 °C. In the equatorial Pacific during La Niña years, most of the prediction Bias values of the TWF (sst, ssw) model were less than 0.8 °C, and the monthly mean Poi_RMSE was less than 1.2 °C. The Poi_RMSE in El Niño and La Niña years was mainly concentrated in the range of 0.4–0.8 °C. The TWF(sst, ssw) model had better prediction performance in El Niño and La Niña years, but the prediction results of the model in La Niña years were better than those in El Niño years. This may have been related to the weak La Niña activity from 2017 to 2018 and the very strong El Niño activity from 2015 to 2016.